Journal of Food Composition and Analysis 35 (2014) 55–60

Original Research Article

Application of multivariate calibration and NIR spectroscopy for thequantification of methylxanthines in yerba mate (Ilex paraguariensis)

Larize Mazur a,*, Patricio Guillermo Peralta-Zamora b, Bogdan Demczuk Jr.a,Rosemary Hoffmann Ribani a

a Graduate Program of Food Engineering, Chemical Engineering Department, Federal University of Parana – UFPR, 81531-980 Curitiba, Parana, Brazilb Graduate Program of Chemical Department, Federal University of Parana – UFPR, 81531-980 Curitiba, Parana, Brazil

A R T I C L E I N F O

Article history:

Received 2 April 2013

Received in revised form 12 March 2014

Accepted 5 April 2014

Keywords:

Yerba mate

Ilex paraguariensis

Methylxanthine

Food composition

NIR spectroscopy

Multivariate calibration

PLS

Non-destructive food analysis

A B S T R A C T

Yerba mate is native to subtropical forests and grows naturally in Brazil, Paraguay and Argentina. Its

consumption has increased due to the health benefits attributed to the presence of phytochemical

compounds, mainly methylxanthines, polyphenols and saponins. Methylxanthines have pharmacologi-

cal properties such as stimulation of the central nervous system, peripheral vasoconstriction, relaxation

of smooth muscle and stimulation of the myocardium. Near infrared (NIR) spectroscopy is a non-

destructive method that provides rapid and simple analysis, facilitating the execution of quality control

operations. The goal of the present study was to evaluate the use of near infrared spectroscopy, combined

with multivariate calibration techniques, to predict the total methylxanthine content in yerba mate.

Several multivariate calibration models were developed using PLS and evaluated by RMSECV and R2

values. The best model was developed using smoothing as pre-processing, the second derivative and

MSC, starting from four latent variables (LV). This method presented R2 and RMSECV values of 0.924 and

0.2, respectively, with good predictive capability during external validation with a percent error of 7.5%.

NIR analysis can be applied to predict the total methylxanthine content in yerba mate.

� 2014 Elsevier Inc. All rights reserved.

Contents lists available at ScienceDirect

Journal of Food Composition and Analysis

jo u rn al ho m epag e: ww w.els evier . c om / lo cat e/ j fc a

1. Introduction

Yerba mate (Ilex paraguariensis) is a plant vastly cultivated andappreciated in South America and is normally consumed in theform of mate, terere and tea (Cardozo et al., 2007; Pagliosa et al.,2010). In Brazil, per capita consumption reaches values close to1.2 kg per year, which is favorable due to the beneficial effects ofyerba mate on human health. In vitro studies have shown thatyerba mate extract contributes to the prevention of cancer,cardiovascular diseases and DNA damage (Menini et al., 2007),mainly due to the presence of polyphenols, saponins andmethylxanthines (Cansian et al., 2008). Among the methyl-xanthines, caffeine, theobromine and theophylline stand out, withthe first two in greater quantities (Clifford and Ramirez-Martinez,1990; Cardozo et al., 2007). Intake of a yerba mate infusionimproves the antioxidant capacity and the resistance of plasmaand LDL particles to ex vivo lipid peroxidation (Silva et al., 2008).

* Corresponding author. Tel.: +55 41 3361 3232; fax: +55 41 3361 3232.

E-mail addresses: larizemazur@yahoo.com.br, larizemazur@hotmail.com

(L. Mazur).

http://dx.doi.org/10.1016/j.jfca.2014.04.005

0889-1575/� 2014 Elsevier Inc. All rights reserved.

Methylxanthines are alkaloids naturally present in yerba matethat have several pharmacological properties, including stimula-tion of the central nervous system, peripheral vasoconstriction,relaxation of smooth muscle and stimulation of the myocardium,as well as diuretic, anti-inflammatory and anti-rheumatic effects(Kikatani et al., 1993; Esmelindro et al., 2002; Lorist and Tops,2003). Caffeine is widely used in the pharmaceutical and cosmeticsindustries (Velasco et al., 2008).

For efficient extraction of methylxanthines from the matrix,several methods have been proposed, including aqueous infusion,sulfuric acid extraction followed by organic solvents (Reginattoet al., 1999) and decoction and extraction with organic solvents(Coelho et al., 2007). Acid extraction by decoction is more effectivefor release of caffeine and theobromine (Gnoatto et al., 2007).

Quantification of methylxanthines by HPLC is widely employedusing isocratic elution systems (Reginatto et al., 1999; Athaydeet al., 2000) or gradient elution (Cardozo et al., 2007). Although theprecision and accuracy of this technique cannot be questioned,practical drawbacks in the form of prolonged analysis time, highconsumption of solvents and the impossibility of carrying outonline analysis encourage the search for new alternative methodsfor analysis. Within this context, attention should be paid to near

L. Mazur et al. / Journal of Food Composition and Analysis 35 (2014) 55–6056

infrared (NIR) spectroscopy, a technique characterized by its speedand reliability, especially when combined with multivariatecalibration tools.

The first studies performed with the use of NIR spectroscopy asan industrial tool were conducted by Karl Norris at the UnitedStates Department of Agriculture in the 1970 (Williams and Norris,2001). NIR spectroscopy is able to provide rapid results andemploys a non-destructive method, without the generation oftoxic byproducts, since it does not use chemical reagents in theanalysis, and requires simple preparation of the sample to beanalyzed (Burns and Ciurczak, 2001; Skoog et al., 2002).

NIR spectroscopy has proven to be an important tool for thequantification of caffeine in coffee (Huck et al., 2005; Rodriguez-Saona et al., 2005), as well as for the determination of caffeine andtheobromine in green tea (Shinija and Mishra, 2009).

This study proposes the use of NIR spectroscopy associated withpartial least squares (PLS) regression for the determination of themethylxanthine content in samples of yerba mate. As far as it waspossible to investigate, only one previous study has described theuse of multivariate tools (principal component analysis; PCA) forthe classification of yerba mate samples by employing the nuclearmagnetic resonance technique (Kim et al., 2010).

2. Materials and methods

2.1. Materials

Samples of yerba mate with high methylxanthine content from25 different plantation lots (about 25 kg each sample), previouslydried in a microwave and ground, were provided by EmbrapaFlorestas, located in Colombo, Parana, Brazil, and all originated inthe region of Ivaı, Parana, Brazil. These samples were packed inpolypropylene bags and stored under refrigeration. Caffeine(�98%) and theobromine (�99%) were purchased from Sigma-Aldrich (St. Louis, MO, USA). Chromatographic grade methanol andsulfuric acid were purchased from Merck (Darmstadt, Germany),while ultrapure water was provided by a Milli-Q system fromMillipore (New Bedford, MA, USA).

2.2. Sample extraction

Yerba mate samples (2 g) were weighed in 100 mL beakers,then 4 mL of sulfuric acid were added and the sample washomogenized and left for 15 min in water. Then, 50 mL of hotwater were added and the samples were left in a bain-marie foranother 15 min. After this time, the samples were filtered throughqualitative filter paper. The extracts were cooled to roomtemperature and neutralized with 40% sodium hydroxide, thentransferred to a 250 mL volumetric flask, where the volume wasfilled with distilled water. The extracts contained in the volumetricflasks were filtered through a cellulose acetate membrane(0.45 mm, Millipore) for subsequent HPLC analysis (Berte et al.,2011). The extraction of the 25 samples was performed induplicate.

2.3. Chromatographic analysis

To perform the chromatographic analysis, was used a highperformance liquid chromatograph Agilent 1200 (Santa Clara, CA,USA), consisting of a quaternary pump (model: G1322A), diodearray detector (model: G1315B), online degasser and dataprocessing system EZ Chrom Elite v. 3.3.1. For all measurements,was used a column Zorbax Eclipse XDB-C18 (4.6 mm � 150 mm,5 mm) (Santa Clara, CA, USA). The mobile phase used in the analysisconsisted of methanol and water (20:80, v/v) at a flow rate of1 mL min�1.

2.4. Near infrared (NIR) spectroscopy

The reflectance spectra of 1 g dry leaves sample of yerba matewere scanned, in the reflectance mode, in the 10,000–4000 cm�1

using a Tensor 37 FTIR spectrometer system (Bruker Optics,Ettlingen, Germany) equipped with an integrative sphere. OPUSsoftware (v. 6.0 Bruker Optics, Ettlingen, Germany) was used forspectral acquisition and instrumental control. Reflectance datawere recorded, at a nominal resolution of 4 cm�1, accumulating 64scans, for NIR spectra.

2.5. Chemometric analysis of data

In the construction of the multivariate calibration models, themethod of partial least squares regression (PLSR) was applied usingthe PLS-toolbox 1.5 package (Eigenvector Research Inc., Manson,WA, USA) working in Matlab 7.0 (Mathworks Inc., Natick, MA,USA). All collected data was organized using Origin 8.01 (North-ampton, MA, USA). From the entire set of 50 spectra, 41 (82%) wereseparated to constitute the calibration set, while 9 (18%) werereserved as the external validation set. The selection of the latterset (nine different spectra) was based on a preliminary evaluationcarried out by principal component analysis (PCA), performed withspectra preprocessed by multiplicative signal correction (MSC) andcentered on the mean and two principal components.

3. Results and discussion

3.1. Chromatographic analysis

The total methylxanthine content found in the 25 samples ofyerba mate ranged from 3.69 to 12.72 mg g�1, whereas theconcentrations of caffeine and theobromine presented resultsbetween 0.001 and 10.11 mg g�1 and 0 02 and 5.03 mg g�1,respectively (Fig. 1). These values are consistent with the resultsreported by several authors in studies of a similar nature (Clifford andRamirez-Martinez, 1990; Cardozo et al., 2007; Berte et al., 2011).

It is important to note that even though the quantification ofthese substrates involved the use of heterogeneous samples andpreliminary extraction processes, the method employed was ableto generate estimates with a relative standard deviation (RSDs)range from 0.1% to 12.87%. The LOD and LOQ were1.6 � 10�3 mg kg�1 and 4.2 � 10�3 mg kg�1 based on the signal-to-noise of 3 and 10, respectively.

3.2. Infrared spectra

The spectra obtained from solid samples of yerba mate areshown in Fig. 2. It is possible to observe good homogeneity in the setof spectra. The signals observed between 5200 and 4200 cm�1 maybe attributed to stretching vibrations of C55O, C55C and C55N bondspresent in methylxanthines, while the bands at 6900 and 8800 cm�1

correspond to methyl groups (CH3) (Shinija and Mishra, 2009).Due to the usual non-specificity of the spectral signal recorded

in the NIR region, characterized by broad, overlapping bands, NIRspectroscopy has been long considered difficult to understand.Improved interpretation has been made possible by the use ofmultivariate techniques, mainly partial least squares (PLS)regression, a technique that allows the construction of calibrationmodels based on practically all spectral information available (Hallet al., 1996; McShane and Cote, 1998).

3.3. Multivariate model development

The main methylxanthines found in mate are caffeine andtheobromine, which are chemical species of great structural

Fig. 1. Box-plot representation of caffeine, theobromine and total methylxanthine concentrations in yerba mate samples (n = 25, duplicate analysis).

L. Mazur et al. / Journal of Food Composition and Analysis 35 (2014) 55–60 57

similarity and possess very similar spectral characteristics.Therefore, multivariate calibration tools may be used to developmodels that aim for the determination of total methylxanthines,resorting to the spectral signal found in the NIR region. Similarprocedures have been adopted by several authors, for instance forthe characterization of coffee and green tea (Paradkar andIrudayaraj, 2002; Chen et al., 2006; Shinija and Mishra, 2009).

Initially, all the spectra (n = 50) underwent a preliminaryexploratory study, using a principal components analysis (PCA)routine. In this study, the spectra were preprocessed by multiplica-tive signal correction (MSC), an algorithm developed for correctingthe effects of scattering on reflectance measurements, and centered

Fig. 2. Infrared spectra of 25 yerb

on the mean, preventing the most distant points from the center ofthe data from having a greater influence than the closest points.

First of all, it is important to emphasize that, as proposed in theprincipal component analysis, the first goal is to compress the datawithout the loss of relevant information. In this case, two principalcomponents (PCs) were used to represent approximately 90% ofthe variance of the spectral data, initially represented byapproximately 1500 transmittance values, recorded in 1500 wavenumber values. Classifying the samples according to the coordi-nates defined by the new set of PCs (scores), it was possible toobserve the differentiation of several samples of yerba mate; thiswas based on 90% of the spectral information (Fig. 3). On the other

a mate samples (duplicate).

Fig. 3. Scores (PC1 � PC2) of 25 yerba mate samples (duplicate).

L. Mazur et al. / Journal of Food Composition and Analysis 35 (2014) 55–6058

hand, it was possible to observe an excellent concordance betweensome duplicates (see samples 16 and 17), while others showedsignificant differences (see samples 22 and 23), probably due to theheterogeneity of the matrix being analyzed. Such differences mustbe considered when assessing the predictive power of multivariatemodels, since prediction errors may not be associated with failuresof the models, but rather to the irreproducibility of spectraacquisition, due to the already mentioned sample heterogeneity.

The selection of the external validation samples (n = 9) wasbased on this preliminary classification, selecting samples from

Fig. 4. Evolution of the coefficient of determination in predicting the calibration set (R2cal

(Preprocessing 1: mean centering, Preprocessing 2: smoothing/first derivative and MSC

each of the four quadrants seen in Fig. 3, giving preference tosamples located in quadrants of higher frequency.

Subsequently, several calibration models were developed,using various numbers of latent variables (LVs) and differenttypes of signal preprocessing, mainly mean centering, smoothing/derivation, multiplicative scatter correction (MSC) and standardnormal deviate (SNV). The cross-validation was used to developthe calibrations validation method. Using the coefficient ofdetermination (R2

cal) as the primary criterion for predictions onthe calibration set (Fig. 4), it was shown that the best predictive

) depending on the number of latent variables and the type of presignal processing

, Preprocessing 3: smoothing/MSC and second derivative).

Fig. 5. Leverage versus studentized residuals considering the developed model containing 4 LVs and preprocessed data with smoothing/MSC and the second derivative.

L. Mazur et al. / Journal of Food Composition and Analysis 35 (2014) 55–60 59

capability for the quantification of the methylxanthine content insamples of yerba mate was obtained with the model developedusing 4 LVs and spectral data preprocessed by smoothing/secondderivative followed by MSC (R2 = 0.924).

An important aspect for the optimization of calibration modelsis to examine the presence of anomalous samples (outliers). Inprocessing by PLS, this verification is performed through theevaluation of leverage, a parameter associated with the influenceof each sample in the modeling, as well as studentized residuals(Fig. 5). The threshold value of the leverage (0.3) was determinedby the ratio 3 VL/n (where VL is the number of latent variables usedin modeling and n is the total number of samples), while thethreshold value for the studentized residuals corresponded to�2.5, taking into account a 95% confidence level. Based on thisassessment (Fig. 5), it was possible to show the absence of anomaliesin the calibration set.

To assess the predictive power of the model, an externalvalidation set consisting of nine samples that did not take part inthe development stage of the model was used. The results (Table 1)indicate a mean error on the order of 7% relative to thechromatographic method used as the reference. It is noteworthythat, even with an average deviation on the order of 4%, thechromatographic method provided differences of up to 17% in theanalysis of duplicates, probably due to the already mentioned

Table 1Application of NIR spectroscopy and multivariate calibration for the measurement

of methylxanthines in yerba mate (Ilex paraguariensis).

Sample Real value (mg g�1) Predicted value (mg g�1) Error (%)

1 8.45 7.28 13.85

2 4.96 4.77 3.83

3 7.84 6.44 17.80

4 7.41 8.37 12.90

5 7.09 6.98 1.55

6 5.91 6.35 7.44

7 4.88 4.59 5.94

8 7.784 7.78 0.05

9 8.13 7.77 4.43

Average error 7.50

heterogeneity of the samples. Thus, obtaining maximum predic-tion errors of this magnitude (samples 1, 3 and 4) is consistent withthe results obtained by chromatography.

Calibration models developed based on NIR spectroscopy forthe quantification of total methylxanthines in green tea, using fourLVs and preprocessing based on the first derivative and normali-zation, prediction errors on the order of 11% were reported, whichis consistent with the results obtained in this work (Shinija andMishra, 2009).

4. Conclusions

The multivariate calibration models based on NIR spectroscopyand developed using a partial least squares (PLS) regression routineallowed for the rapid determination of the methylxanthine contentin samples of yerba mate, giving results consistent with thoseprovided by the reference chromatographic technique. Addition-ally, it is important to note that the combination of NIRspectroscopy with multivariate analysis tools represents apromising alternative for the quality control of this type ofproduct, not only because of the reliability of these techniques, butalso in terms of characteristics such as speed as well as their non-destructive and non-polluting nature (no residue generation) andthe possibility of implementing systems for online analysis. Therapid control of the methylxanthines content in yerba mate in theindustrial processing allows the selection products suitable forspecific groups of consumers and thus meeting their needs andexpectations on the final product.

References

Athayde, M.L., Coelho, G.C., Schenkel, E.P., 2000. Caffeine and theobromine inepicuticular wax of Ilex paraguariensis A. St.-Hil. Phytochemistry 55, 853–857.

Berte, K.A.S., Beux, M.R., Spada, P.K.W.D., Salvador, M., Hoffmann-Ribani, R., 2011.Chemical composition and antioxidant activity of yerba-mate (Ilex paraguar-iensis A. St.-Hil., Aquifoliaceae) extract as obtained by spray drying. Journal ofAgricultural and Food Chemistry 59, 5523–5552.

Burns, D.A., Ciurczak, E.W., 2001. Handbook of Near-Infrared Analysis. MarcelDekker Inc., New York.

Cansian, R.L., Mossi, A.J., Mazutti, M., Oliveira, J.V., Paroul, N., Dariva, C.,Echeverrigaray, S., 2008. Semi-volatile compounds variation among Brazilian

L. Mazur et al. / Journal of Food Composition and Analysis 35 (2014) 55–6060

populations of Ilex paraguariensis St. Hil. Brazilian Archives of Biology andTechnology 51, 175–181.

Cardozo Junior, E.L., Ferrarese-Filho, O., Cardozo Filho, L., Ferrarese, M.L.L., Dona-duzzi, C.M., Sturion, J.A., 2007. Methylxanthines and phenolic compounds inmate (Ilex paraguariensis St. Hil.) progenies grown in Brazil. Journal of FoodComposition and Analysis 20, 553–558.

Chen, Q., Zhao, J., Zhang, H., Wang, X., 2006. Feasibility study on qualitative andquantitative analysis in tea by near infrared spectroscopy with multivariatecalibration. Analytica Chimica Acta 572, 77–84.

Clifford, M.N., Ramirez-Martinez, J.R., 1990. Chlorogenic acids and purine alkaloidscontents of mate (Ilex paraguariensis) leaf and beverage. Food Chemistry 35,13–21.

Coelho, G.C., Rachwal, M.F.G., Dedecek, R.A., Gustavo, R., Curcio, G.R., Nietsche, K.,Schenkel, E.P., 2007. Effect of light intensity on methylxanthine contents of Ilexparaguariensis A. St. Hil. Biochemical Systematics and Ecology 35, 75–80.

Esmelindro, M.C., Toniazzo, G., Waczuk, A., Dariva, C., Oliveira, D., 2002. Caracter-izacao fısico-quımica da erva-mate: Influencia das etapas do processamentoindustrial. Ciencia e Tecnologia de Alimentos 22, 193–204.

Gnoatto, S.C.B., Bassani, V.L., Coelho, G.C., Schenkel, E.P., 2007. Influencia do metodode extracao nos teores de metilxantinas em erva-mate (Ilex paraguariensis A. St.-Hil., Aquifoliaceae). Quimica Nova 30, 304–307.

Hall, J.W., McNeil, B., Rollins, M.J., Draper, I., Thompson, B.G., Macaloney, G., 1996.Near-infrared spectroscopic determination of acetate, ammonium, biomass,and glycerol in an industrial Escherichia coli fermentation. Applied Spectroscopy50, 102–108.

Huck, C.W., Uggeabichler, W.G., Bonn, G.K., 2005. Analysis of caffeine theobromineand theophylline in coffee by NIR spectroscopy compared to HPLC coupled tomass spectrometry. Analytica Chimica Acta 538, 195–203.

Kikatani, T., Watanabe, Y., Shibuya, T., 1993. Different effects of methylxanthines oncentral serotonergic postsynaptic neurons in a mouse behavioral model. Phar-macology Biochemistry and Behavior 44, 457–461.

Kim, H.K., Saifullah Khan, S., Wilson, E.G., Kricun, S.D., Meissner, A., Goraler, S.,Deelder, A.M., Choi, Y.H., Verpoorte, R., 2010. Metabolic classification of SouthAmerican Ilex species by NMR-based metabolomics. Phytochemistry 71, 773–784.

Lorist, M.M., Tops, M., 2003. Caffeine, fatigue and cognition. Brain and Cognition 53,82–94.

McShane, M.J., Cote, G.L., 1998. Near-infrared spectroscopy for determination ofglucose lactate and ammonia in cell culture media. Applied Spectroscopy 52,1073–1078.

Menini, T., Heck, C., Schulze, J., Mejia, E., Gugliucci, A., 2007. Protective action of Ilexparaguariensis extract against free radical inactivation of Paraoxonase-1 inHight-Density Lipoprotein. Planta Medica 73, 1141–1147.

Pagliosa, C.M., Vieira, M.A., Podesta, R., Maraschin, M., Zeni, A.L.B., Amante, E.R.,Amboni, R.D.M.C., 2010. Methylxanthines, phenolic composition, and antioxi-dant activity of bark from residues from mate tree harvesting (Ilex paraguar-iensis A. St. Hil.). Food Chemistry 122, 173–178.

Paradkar, M.J., Irudayaraj, J., 2002. A rapid FTNIR spectroscopy method for estima-tion of caffeine in soft drinks and total methylxanthines in tea and coffee.Journal of Food Science 67, 2507–2511.

Reginatto, F.H., Athayde, M.L., Gosmann, G., Schenkel, E.P., 1999. Methylxanthinesaccumulation in Ilex species – caffeine and theobromine in erva mate (Ilexparaguariensis) and other Ilex species. Journal of the Brazilian Chemical Society10, 443–446.

Rodriguez-Saona, L.E., Fry, F.S., Calvery, E., 2005. Use of Fourier transform nearinfrared spectroscopy rapid quantification of castor bean meal in a selection offlour based product. Journal of Agricultural and Food Chemistry 48, 5169–5177.

Shinija, V.R., Mishra, H.N., 2009. FT-NIR spectroscopy for caffeine estimation ininstant green tea powder and granules. Food Science and Technology 42, 998–1002.

Silva, E.L., Neiva, T.J.C., Shirai, M., Terao, J., Abdalla, D.S.P., 2008. Acute ingestion ofyerba mate infusion (Ilex paraguariensis) inhibits plasma and lipoprotein oxi-dation. Food Research International 41, 973–979.

Skoog, D.A., Holler, F.J., Nieman, T., 2002. Principios de analises instrumental.Bookman, Porto Alegre.

Velasco, M.V.R., Tano, C.T., Machado-Santelli, G.M., Consiglieri, V.O., Kaneko, T.M.,Baby, A.R., 2008. Effects of caffeine and siloxanetriol alginate caffeine, asanticellulite agents, on fatty tissue: histological evaluation. Journal of CosmeticDermatology 7, 23–29.

Williams, P., Norris, K.H., 2001. Variables affecting near infrared spectroscopicanalysis. In: Williams, P., Norris, K.H. (Eds.), Near Infrared Technology in theAgriculture and Food Industries. 2nd ed. The American Association of CerealChemists, St. Paul, MN, pp. 171–185.